Generative AI in call centres creates value when it improves resolution, reduces avoidable effort, and strengthens knowledge flow without weakening trust, privacy, or accountability. In 2026, the best deployments stay narrow at first, use grounded knowledge, keep humans in sensitive moments, and measure outcomes like first contact resolution, repeat contact, and compliance rather than bot activity alone.¹˒²˒³ (IDEAS/RePEc)
What is generative AI in call centres?
Generative AI in call centres means using large language models and related tools to create, transform, or summarise language inside customer service work. That includes drafting responses, summarising calls, retrieving and reshaping knowledge, classifying intent, supporting agents in real time, and powering self-service conversations. It is broader than a chatbot. It touches workflow, quality, knowledge, and operating control.³˒⁷ (NIST Publications)
The difference matters. Traditional automation follows a fixed path. Generative AI works with messy language, mixed intent, and incomplete context. That makes it useful in service environments where customer requests arrive through voice, chat, email, and case notes rather than neat form fields. But it also makes accuracy, explainability, and governance harder.¹˒³ (IDEAS/RePEc)
Why are call centres investing now?
The pressure is practical. Contact centres still face high interaction volumes, rising customer expectations, and constant pressure to improve cost and service quality together. McKinsey noted in 2025 that AI-driven solutions can already solve simple transactional issues through virtual assistants and knowledge-based service flows, shifting some work away from live channels.⁹ That is the reward side of the case. (McKinsey & Company)
But the 2026 context changes the decision. The technology is more capable, yet governance expectations are tighter. The OECD’s Due Diligence Guidance for Responsible AI, published on 19 February 2026, tells enterprises to address adverse impacts proactively and monitor AI use continuously.⁴ In Australia, OAIC guidance says the Privacy Act applies to uses of AI involving personal information, and APRA’s CPS 230 is now in force for regulated entities.⁵˒⁶ (OECD)
How do the rewards actually show up?
The clearest rewards sit in language-heavy work. Generative AI can cut search time, reduce after-call work, speed up drafting, improve consistency, and help agents handle more variation without relying on memory alone. Research on GenAI-enabled customer service points to lower service cost, faster interactions, and more scalable personalisation as genuine opportunities.¹ Meanwhile, chatbot adoption research from 2025 found that GenAI can lift perceived usefulness, human-likeness, and familiarity.² (IDEAS/RePEc)
There is also an operating reward that gets less attention. Knowledge moves faster. Good GenAI setups can turn repeated customer questions, agent workarounds, and call summaries into updated guidance and clearer service content. In live operations, that often matters more than flashy conversation quality because better knowledge improves both self-service and agent-assisted channels.⁷ (sciencedirect.com)
What is the difference between assistive and autonomous use?
Assistive GenAI supports the human. It drafts, summarises, suggests, and retrieves, but the person still decides what to send or do. Autonomous or semi-autonomous GenAI takes a larger role by responding directly, routing work, or completing bounded workflow steps. Most contact centres should begin with the first model. It is safer, easier to govern, and more likely to build workforce trust.³˒⁷ (NIST Publications)
That is not caution for its own sake. It reflects the real trade-off. The more authority the model has, the more exposed the organisation becomes to incorrect answers, privacy leakage, biased treatment, and service failures at scale. NIST’s Generative AI Profile is explicit that trustworthiness risks for GenAI include confabulation, information integrity failures, privacy risks, and broader system vulnerabilities.³ (NIST Publications)
Which use cases make sense first?
Start with tasks where the language burden is high and the downside of error is manageable. Good first candidates are call summarisation, after-call notes, knowledge retrieval, draft email and chat responses, complaint intake triage, and agent-assist prompts. These use cases usually save time without handing over too much discretion.¹˒³˒⁷ (IDEAS/RePEc)
In that stage, a knowledge-led product such as Knowledge Quest is the right kind of first link because it focuses on real-time, accurate answers, knowledge health, and faster updates rather than free-form automation alone. That matters because weak knowledge underneath GenAI usually scales inconsistency faster, not slower. (Customer Science)
What should stay human-led?
High-emotion and high-discretion work should stay human-led. Complaints, service recovery, vulnerability, hardship, bereavement, disputes, and complex exception handling all depend on judgment and emotional skill as much as language generation. Research in 2025 found that customers perceive service providers as less customer-oriented when recovery is handled by voice AI rather than a human, especially in feeling-heavy moments.⁸ (sciencedirect.com)
This is where many deployments go wrong. Leaders see that GenAI writes well and assume it should lead the whole interaction. But language fluency is not empathy, and speed is not trust. The better design is blended. Let the model prepare context, summarise facts, and suggest language. Let the human own the emotionally or commercially sensitive decision.¹˒⁸ (IDEAS/RePEc)
What are the real risks?
The first risk is accuracy. GenAI can produce persuasive but unsupported answers. The second is privacy. Once personal information moves through commercial AI products, Australian privacy obligations apply. The third is operational risk. If GenAI sits inside critical service workflows, poor resilience or poor vendor control becomes an executive issue, not just a technical one.³˒⁵˒⁶ (NIST Publications)
There are softer risks too. Customers may find the service helpful yet still feel watched, pushed, or emotionally short-changed. The 2024 Business Horizons article on GenAI customer service framed this well through six paradoxes, including lower cost yet higher price, higher quality yet less empathy, and personalised yet intrusive.¹ Those tensions are not abstract. They show up when a contact centre uses GenAI in the wrong moment or with weak escalation design. (IDEAS/RePEc)
How should you measure rewards against risk?
Do not measure deployment by bot volume or model usage. Measure the service system. The reward side should include first contact resolution, repeat contact within seven days, handle time, after-call work, knowledge-search time, draft acceptance rate, complaint volume, and employee confidence. The risk side should include hallucination incidence, override rate, privacy incidents, fallback frequency, and policy compliance defects.³˒⁴ (NIST Publications)
This is usually where outside support matters. CX Consulting and Professional Services is relevant in the measurement and rollout phase because the hard part is rarely the model alone. It is sequencing use cases, setting controls, redesigning workflows, and proving value in normal governance forums. (Customer Science)
What should happen next?
Pick three use cases, not thirty. One knowledge use case. One agent productivity use case. One customer-facing use case with a clear fallback path. Define the knowledge source, the action boundary, the human escalation rule, and the scorecard before go-live. Then test under real load with a real service team.³˒⁴˒⁶ (NIST Publications)
The point is simple. A good GenAI contact-centre deployment is not a software launch. It is an operating-model change. The organisations that get the rewards will be the ones that treat it that way.
FAQ
What is the best first use of generative AI in call centres?
The best first use is usually knowledge retrieval or call summarisation because both reduce effort, both are measurable, and neither requires the model to make a sensitive customer decision on its own.³˒⁷ (NIST Publications)
Does GenAI improve customer trust automatically?
No. Research shows GenAI can increase usefulness and familiarity, but trust does not automatically rise, and privacy concerns can increase.² (sciencedirect.com)
Should GenAI handle complaints and recovery?
Only in a support role at first. It can gather facts, draft explanations, and suggest next actions, but humans should handle emotionally loaded or discretionary recovery work.¹˒⁸ (IDEAS/RePEc)
What are the main Australian governance issues?
Privacy, operational resilience, vendor risk, and accountability are the main ones. OAIC guidance applies when personal information is used with commercial AI products, and APRA-regulated entities also need to account for CPS 230 requirements.⁵˒⁶ (OAIC)
What usually blocks a good deployment?
Weak knowledge, unclear ownership, poor escalation design, and shallow measurement block good deployment more often than the model itself.³˒⁷ (NIST Publications)
What helps keep AI-written service messages clear and brand-safe?
CommScore.AI is relevant where teams need written responses checked for clarity, tone, and brand alignment before poor communication quality starts creating extra contact and distrust. (Customer Science)
Evidentiary Layer
The balance of evidence is clear enough to act on. Generative AI in call centres can improve speed, consistency, and knowledge handling. But the same evidence says rewards are conditional. They depend on grounded knowledge, bounded scope, human oversight, and governance that treats privacy, resilience, and trust as operating requirements rather than legal fine print.¹˒²˒³˒⁴˒⁵˒⁶˒⁸ (IDEAS/RePEc)
Sources
-
Ferraro, C., Demsar, V., Sands, S., et al. The paradoxes of generative AI-enabled customer service. Business Horizons, 2024. DOI: 10.1016/j.bushor.2024.04.014.
-
Arce-Urriza, M., Cebollada, J., Tarifa-Fernández, J. From familiarity to acceptance: The impact of generative AI on chatbot adoption. Journal of Retailing and Consumer Services, 2025. DOI: 10.1016/j.jretconser.2024.104089.
-
NIST. Artificial Intelligence Risk Management Framework: Generative AI Profile. NIST AI 600-1, 2024.
-
OECD. OECD Due Diligence Guidance for Responsible AI. 19 February 2026.
-
OAIC. Guidance on privacy and the use of commercially available AI products. 21 October 2024.
-
APRA. Prudential Standard CPS 230 Operational Risk Management. In force from 1 July 2025.
-
Ledro, C., Nosella, A., Vinelli, A. Artificial intelligence in customer relationship management: A systematic framework for successful integration. Journal of Business Research, 2025. DOI: 10.1016/j.jbusres.2025.115214.
-
Carrilho, M. G., Wagner, R., Pinto, D. C., et al. The role of empathy in voice-driven AI for service recovery. Journal of Business Research, 2025. DOI: 10.1016/j.jbusres.2025.115703.
-
McKinsey & Company. The contact center crossroads: Finding the right mix of humans and AI. 19 March 2025.